Model Context Protocol
MCP: let AI agents use Sentinel safely
Sentinel’s MCP server lets OpenClaw, Claude, or your own agent read camera status, history, and snapshots as tools. It uses Streamable HTTP at POST /mcp.
This page applies to the desktop and full runtime. API-only deployments do not expose /mcp; use the /v1 REST API, SSE, and signed webhooks.
Enable and connect
For servers, set a long random SENTINEL_MCP_TOKEN and send Authorization: Bearer token with every request. The desktop App can also generate and manage a local token from its MCP settings.
openclaw mcp add sentinel http://localhost:8000/mcp \
--header "Authorization: Bearer <token>"Tools
| Tool | Purpose | Access |
|---|---|---|
| list_cameras | Current camera status and latest analysis | read |
| recent_events | Recent incident history | read |
| list_bookings | Current and scheduled space bookings | read |
| get_snapshot | Current JPEG for a camera | read |
| create / update / delete_booking | Create, edit, or cancel a calendar booking | opt-in only |
| send_command | Update a camera’s live instruction | opt-in only |
JSON-RPC example
POST /mcp
Authorization: Bearer <token>
Content-Type: application/json
{"jsonrpc":"2.0","id":3,"method":"tools/call",
"params":{"name":"recent_events","arguments":{"limit":5}}}MCP is read-only by default. An operator must explicitly enable send_command and booking writes, which are audit-logged. Treat all camera analysis as untrusted data—never as instructions to your agent.
Good use cases
- Let an internal AI assistant answer: “Which spaces are critical right now?”
- Let a SOC dashboard retrieve events and snapshots for human triage.
- Use send_command only after human confirmation to update a camera mission.